JP3652086B2 - Speed reading support device - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention relates to the entire Japanese document processing, and is used for an interface for speed reading support of an electronic document and information retrieval.
[0002]
[Prior art]
Conventionally, when a summary sentence candidate for a sentence S is determined by a computer, the possibility that the sentence s in the sentence S becomes a summary sentence is calculated using a scale called probability P or TFIDF. In general, priorities are assigned to candidates for summary sentences. For example, Julian Kupiec, Jan Pedersen, and Francine Chen. 1995. “A Trainable Document Summarizer” In Proceedings of the Fourteenth Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 68-73, Seattle, USA. The probability P that a certain sentence s in the sentence S is selected as a summary sentence is calculated using the formula of 1).
P (s inX | F1, F2, F3,... Fk) (Equation 1)
Here, X represents a set of summary sentences, F1, F2,..., Fk represent features such as sentence length, presence / absence of clue words, position within a paragraph, and the like. Then, the importance of each sentence as a summary sentence is determined based on the formula (Equation 1), and sentences corresponding to the top 25% of the values are presented to the user as sentence summaries.
[0003]
On the other hand, Klaus Zechner. 1996. “Fast Generation of Abstracts from General Domain Text Corpora by Extracting Relevant Sentences” In Proceedings of the 16th International Conference on Computational Linguistics, pages 986-989. Copenhagen, Denmark. Calculate the scale according to Equation 2) to determine the importance of the sentence s.
TFIDF (w, s) = TF (w, s) × log (N / n (w)) (Equation 2)
Here, w is a word that appears in a specific sentence s, TF (w) is the frequency of the word w in the sentence, N is the total number of sentences s in the sentence S, and n (w) is the sentence s in which the word w appears. It is the total number. The importance Q (s) of the sentence s is defined by (Equation 3).
Q (s) = Σ TFIDF (w, s) (Equation 3)
That is, for all words appearing in the sentence s, the TFIDF values are obtained and the sum is taken as the importance of the sentence s. Then, a sentence having a higher Q (s) value is presented to the user as a summary candidate. The Zechner method described above generally has poor summarization accuracy because it cannot be tuned specifically for each field.
[0004]
[Problems to be solved by the invention]
In these conventional methods, the evaluation of the correctness as a summary sentence is put aside, and the selected summary sentence is displayed by itself, so even if several upper sentences are displayed as a summary sentence, it is selected. Since each summary sentence is not originally related, the connection before and after those sentences deteriorates and becomes very difficult to read. Furthermore, since some selected sentences are simply enumerated, it is difficult to grasp the meaning of the original sentence, and it is difficult to determine whether the sentence is important for the user.
[0005]
An object of the present invention is to solve the problems of the conventional methods described above.
[0006]
[Means for Solving the Problems]
Therefore, in the present invention, the input sentence is subjected to feature analysis for each sentence in the sentence according to a predetermined rule to determine whether or not the sentence is a summary sentence. Each paragraph and first sentence of the input sentence is presented to the user in a highlighted color different from the summary sentence.
[0007]
DETAILED DESCRIPTION OF THE INVENTION
First, a well-known decision tree construction method called C4.5 is used as a summary sentence selection method. If this method is followed, the sentence will be coded based on several features. In the present invention, first, each sentence s in the sentence S is represented by (1) sentence type, (2) position in the sentence, (3) similarity to the headline, (4) TFIDF in the sentence, (5) attitude expression. And (6) the number of characters in the sentence, and (7) the position within the paragraph.
“Sentence type” indicates to which type the sentence belongs, such as a news report, an editorial, or an essay.
[0008]
“Position in sentence” indicates the position at which the sentence appears in the entire sentence. For example, if the total number of sentences s in the sentence S is 10, and the sentence appears in the first sentence, the position of the sentence is represented as 0/10 = 0.
[0009]
The “similarity with the headline” is determined by the following (Equation 4).
SIMM (t, s) = Σ NF (w, s) × IDF (w) (Equation 4)
Here, t represents a sentence head and s represents a sentence. For the noun w appearing in t, its NF value and IDF value are obtained, and the sum is taken as the similarity to the heading. NF (w, s) is defined as (Equation 5).
NF (w, s) = F (w, s) / MAX_F (s) (Equation 5)
Here, F (w, s) is the frequency of w in s, and MAX_F (s) is the frequency of the noun having the highest frequency among the nouns appearing in the sentence s. IDF (w) is defined as (Equation 6).
IDF (w) = log (N / DF (w)) / logN (Equation 6)
Here, DF (w) is the total number of sentences in which w appears. N is the total number of sentences s in the sentence S.
[0010]
“In-text TFIDF” is a value determined by (Equation 7).
D (s) = Σ NF (w, s) × IDF (w) (Expression 7)
Here, w is a noun that appears in the sentence s. NF (w, s) and IDF (w) follow the above definition.
[0011]
“Presence / absence of attitude expression” is used to indicate whether or not there is an expression indicating the author's attitude at the end of the sentence. Here, as expressions expressing the attitude of the author, expressions such as “˜important”, “˜necessary”, “˜ka”, “˜yo”, “˜ne” are considered.
[0012]
The “number of characters in the sentence” indicates the number of characters in the sentence s.
[0013]
“Position within a paragraph” indicates the position within a paragraph of a sentence s by the number of sentences preceding the sentence s / the total number of sentences in the paragraph, as in the above-described “position within a sentence”.
[0014]
The decision tree is structured by characterizing each sentence of the sentence with respect to the above-mentioned attributes and further learning using the data with summary judgment information for each field. The decision tree construction method follows “C4.5” by Quinlan. The decision tree construction method C4.5 is a well-known method, but the outline is as follows.
[0015]
In C4.5, the issue is how to model (generalize) the classification of database entries. For example, suppose a company's recruitment database is as follows.
[0016]
Gender Age Marriage Educational background Car adoption
Female 23 Married High school Yes ○
Male 30 Single University None ○
Female 45 Married High school Yes ○
Male 60 Married University None ×
Classification modeling is to find a pattern of conditions for adoption / non-recruitment from this data and to predict whether or not a person will be employed by this company. By the way, the “recruited” items in the database are classified, and the other items are called attributes. Each entry is called a case. In C4.5, while observing case attribute information, cases with similar attribute values are grouped and classified.
[0017]
For example, in the above example, the following classification model is possible.
What's your gender?
Female
Men What is a car license?
Yes → Adopt
None → Not adopted
In other words, it is possible to generalize “all women are hired” and “male who have a car license”. In actual situations, there is a problem of which attribute is used as a branching condition. However, in C4.5, attributes are selected using a statistical measure called gain ratio.
[0018]
Next, a speed reading support operation is performed using the obtained decision tree. The operation includes the following procedures.
(1) Call the sentence S you want to read quickly on the screen. (2) Use the decision tree to classify each displayed sentence s as a summary sentence. (3) Classify as a summary sentence. (4) Finally, the first sentence of each paragraph is displayed in an emphasized color. (4) Finally, the first sentence of each paragraph is displayed in an emphasized color.
[0019]
As described above, when constructing a decision tree for each field, a summary sentence specialized for a specific field can be generated. In addition, by presenting the summary text to the user in the highlighted color and other text in the background color, the display of the text will be conspicuous, and since the summary text is displayed as it is in the text, The context of the text is saved and can be easily referenced as needed, making it easier to read the summary. Furthermore, by displaying the first sentence of each paragraph together with the summary, the outline of the contents can be understood. Hereinafter, more specific embodiments will be described with reference to the drawings.
[0020]
Example 1
FIG. 1 is a block diagram showing the concept of data processing of the speed reading support method according to the present invention.
[0021]
In FIG. 1, reference numeral 1 denotes an input step that takes in a document (article) to be read at high speed or takes in a document (article) input in advance in a database. Here, it is assumed that the article shown in FIG. 2 is input. 2 is a genre information acquisition step, in which a genre corresponding to the content of the input article is determined and output. The genre determination procedure is as follows. First, a keyword indicating a genre is searched for in the input article. If a keyword representing a genre cannot be found, the user is requested to input genre information. However, in the article of FIG. 2, since the heading portion has the keyword “(editorial)”, the genre of the article is determined to be editorial.
[0022]
Next, the acquired genre information and article are sent to decision tree selection step 3. Here, the article corresponding to the genre of the article is selected from a database of decision trees prepared in advance based on the genre information. The decision tree is constructed according to C4.5 described above. Since the genre of the article is editorial, the editorial decision tree database is selected. In this description, it is assumed that the decision tree shown in FIG. 4 is selected.
Next, the process proceeds to feature extraction step 4. Here, each sentence appearing in the text excluding the headline of the article is subjected to morphological analysis processing, and then feature extraction is performed. The morphological analysis follows the procedure of, for example, [Sakurai et al., Design and Evaluation of Morphological Analysis Program ANIMA, Information Processing Society of Japan 54th Annual Conference, 1997]. Features to be extracted are (1) sentence type, (2) position in sentence, (3) similarity to headline, (4) TFIDF value in sentence, (5) presence / absence of attitude expression, (6) number of characters in sentence (7) Seven positions in the paragraph.
[0023]
Extraction includes the following procedures. Let s be an arbitrary sentence in the text. In the case of newspaper articles provided electronically, classification information such as general articles, essays, and editorials is usually attached to the sentence S. The sentence type of sentence s follows the classification information. When there is no classification information, the user distinguishes general articles, essays, and editorials, and determines the type. The position in the sentence is given by the ratio of the sentence appearing from the beginning of the text to immediately before the sentence s in the entire sentence, that is, D (s) / N. Here, D (s) is the number of sentences from the beginning of the text to immediately before the sentence s, and N is the total number of sentences in the sentence S. As for the similarity to the headline, if the headline of the sentence is T, SIMM (T, s) is calculated according to the above-described (Equation 4), and the value is set as the similarity. The in-sentence TFIDF calculates NF (s, w) × IDF (w) for each noun w appearing in the sentence s, and uses the sum as a value. (However, the noun w is extracted by morphological analysis.) The presence / absence of the attitude expression is specific to the sentence s "~ important", "~ necessary", "~ ka", "~ yo" ”, Etc. (the end form if it is used). If there is nothing here, 1 is set, 2 is set for attitude verbs such as “important” and “necessary”, and 3 is set for final particles such as “ka”, “yo”, and “ne”. The number of characters in the sentence is the number of characters in the sentence s. The position in the paragraph is given as PD (s) / N (P) of the sentence s. Here, PD (s) is the number of sentences appearing from the beginning of the paragraph to immediately before the sentence s, and N (P) is the total number of sentences in the paragraph.
[0024]
FIG. 3 shows the result of the feature extraction performed on each sentence of the article in FIG. 2 by the above procedure. In this embodiment, there are seven sentences excluding the heading, sentence 1 is the first sentence of the body, sentence 2 is the second sentence, sentence 3 is the third sentence,..., Sentence 7 is the seventh sentence. Corresponds to condition. Since all the sentences in the main text are part of the editorial, the sentence type is all “Editorial”. (1) type of sentence extracted, (2) position in sentence, (3) similarity to headline, (4) TFIDF value in sentence, (5) presence / absence of attitude expression, (6) number of characters in sentence, (7) The seven features of the position in the paragraph are as shown in the figure.
[0025]
In the next step 5, it is determined whether the sentence is a summary sentence based on the features extracted from the sentence and the selected decision tree. In the following, the actual determination work will be examined in detail for sentence 1 to sentence 7. Assume that the decision tree is shown in FIG.
[0026]
Since sentence 1 has a similarity to the headline of 0.679, the process proceeds to N10. Furthermore, since the degree of similarity is 1.181 or less, N12 is passed. Next, since TFIDF is 9.449, it passes through N14. Next, since the number of characters is 41, N16 is passed. Next, since TFIDF is 9.449, it passes through N18 and is finally determined as a non-summary sentence.
[0027]
Since sentence 2 has a similarity to the headline of 0.263, it proceeds to N1. However, since the position in the sentence is 0, it is determined as a non-summary sentence through N2.
[0028]
Sentence 3 passes N10 and N12 because the similarity to the headline is 0.762. Furthermore, since the TFIDF value is 4.893, N14 is passed. In the next step, since the number of characters is 70, it passes through N15 and is determined as a summary sentence.
[0029]
Sentence 4 passes N1 because the similarity to the headline is 0.263. However, since the position in the sentence is 0.071, it passes through N2 and is determined as a non-summary sentence.
[0030]
Sentence 5 also passes N1 because the similarity to the headline is 0.263. Since the position in the sentence is 0.095, similarly to sentence 4, it passes through N2 and is determined to be a non-summary sentence.
[0031]
Sentence 6 passes N1 because the similarity to the headline is 0. Since the position in the sentence is 0.119, it is determined as a non-summary sentence through N2 as in the case of sentences 4 and 5.
[0032]
Sentence 7 also passes through N1 because the similarity to the headline is 0, and since the position in the sentence is 0.143, it passes through N2 and is determined to be a non-summary sentence, as in sentences 4-sentence 6. .
[0033]
In the next step 6, according to the determination result of the summary sentence, the sentence determined to be a summary sentence is displayed in an emphasis color, and the sentence determined to be a non-summary sentence is displayed in a background color. Further, in order to add an accent to the display, the first sentence of each paragraph of the article is highlighted with an emphasis color different from that of the summary sentence, and the speed reading support process is terminated.
[0034]
By referring to the corresponding decision tree for texts other than editorials, essays, news reports, etc., and performing processing equivalent to the above, it is possible to support speed reading for texts in other genres.
[0035]
FIG. 5 is a diagram showing a specific processing flow of the above-described processing. In the example of FIG. 5, the acquisition of the sentence type information is performed only once for the sentence that is the target of the fast reading support. On the other hand, the feature extraction of each sentence is performed for each sentence as the unread sentence for each sentence, and the process ends when there is no unprocessed sentence. . The processing flow of FIG. 5 can be easily understood by reading it with reference to the above description, and therefore, description with reference numerals attached to the drawing is omitted. The feature table shown in FIG. 3 described above is a list of features extracted by the processing flow of FIG. 5 for explanation.
[0036]
FIG. 6 shows the display state of the article reflecting the above-described summary sentence and non-summary sentence determination results. Since the figure is not colored, the underlined solid lines are added to the highlighted colors, and the dotted underline is added to the first sentence of each paragraph of the article.
[0037]
In the above-described embodiment, there is a headline in the article, and using this, genre acquisition and similarity evaluation were performed very easily. However, when there is no headline, the similarity of the headline in FIG. Data is lost and the paths N11 and N12 in FIG. 4 are lost, but there is no substantial problem.
[0038]
Of course, if an article is too long to fit on one screen, you can scroll to see the content.
[0039]
Example 2
Next, an embodiment of a newspaper speed reading support device that combines article search support and the above-described speed reading support method will be described.
[0040]
FIG. 7 is a diagram showing a summary of the flow of signal processing for this purpose. Reference numeral 71 denotes a search condition input step for inputting a search condition for an article that the user wants to read. Search conditions can be set arbitrarily, but keywords are common and easy to use. 72 is an article search step, in which article information is fetched from an arbitrary database, and an article meeting the above-described search conditions is searched. Reference numeral 73 denotes a search result display step, which displays articles that meet the search conditions together with the degree of coincidence with the conditions, for example. 74 is an article designation step, in which the user selects an article that he / she wants to read with reference to the degree of coincidence with the condition, for example. 75 is a speed reading support instruction step, which is a step for requesting speed reading support for an article selected by the user for reading. Steps 76 to 80 are speed reading support steps as apparent from FIG. 1, and the speed reading support processing is performed for the article selected by the user in the article specifying step 74.
[0041]
FIG. 8 is a diagram illustrating an example of a hardware configuration for executing this processing. In FIG. 8, reference numeral 801 denotes output means, which means a printer or the like here. Reference numeral 802 denotes a CPU, which executes processing according to a program described later. Reference numeral 803 denotes input means such as a keyboard and a mouse. Reference numeral 804 denotes a system bus. Reference numeral 812 denotes display means, which is a so-called display such as CTP. Reference numeral 809 denotes program holding means, for example, a hard disk is used. The program holding means 809 includes search, speed reading support interface operation program 805, morphological analysis program 806, decision tree generation program 807, feature extraction program 816, decision tree operation program 808, search program 809, similarity calculation program 8091, document ranking. A program 8092, an important sentence display program 810, and a synopsis display program 811 are stored. Reference numeral 813 denotes a work area of the memory. Reference numeral 814 denotes a decision tree database. A document database 815 stores, for example, newspaper articles to be searched. Each means and database are coupled via a system bus.
[0042]
First, when the user wants to read a particular article of interest in the newspaper article, he activates the device to search and activate the fast-reading support interface operating program 805, and displays the search interface input screen of the display means 812. Enter the search keyword. This corresponds to step 71 in FIG. Next, the search program 809 is executed for the input keyword (FIG. 7—Step 72).
[0043]
The search program calculates the similarity D between the text stored in the document database 815 and the input keyword according to (Equation 8).
[0044]
D (q, d) = Σ TF (w, d) × IDF (w) (Equation 8)
Here, q is a list of keywords, d is a document, and is expressed as a list of noun words (excluding duplicates) appearing therein. w represents an element (word) of the list q. TF (w, d) is calculated by the frequency of w in the document d, and IDF (w) is calculated for the entire article stored in the document database 815 according to (Equation 9).
[0045]
IDF (w) = log (N / DF (w)) (Equation 9)
Here, N is the total number of articles in the database, and DF (w) is the total number of articles in which the word w appears even once. However, noun extraction from the document is performed by executing the morphological analysis program 806. The specific method uses the morphological analysis program described in [Asai et al., Design and Evaluation of Morpheme Analysis Program ANIMA, 1997] described above. In this way, the similarity D is obtained for all articles in the document database 815, five articles are selected from those having a high value, and the selection result is displayed to the user via the output screen of the search interface of the display means 812. Present (step 73 in FIG. 7). An example of the output screen adopted here is shown in FIG. In FIG. 9, 91 indicates a headline, and 92 is a switch for turning on the display of the article text. When this switch is clicked, the contents of the corresponding article from the document database are displayed on the display means (FIG. 7-). Step 74).
[0046]
FIG. 10 shows an example of a screen when the user turns on the switch 92 here. In this example, the content of the article is the same as that described with reference to FIG. 2, but in FIG. 10, the user is presented with a selection switch 1001 as to whether or not to support fast reading along with the text (FIG. 7—Step 75) . Here, when the user selects the rapid reading support, the morphological analysis program 806 is executed, morphological analysis is performed for each sentence of the displayed text, and genre information acquisition (FIG. 7-step 76) and decision tree selection are performed (FIG. 7). Along with step 77), the processing result is passed to the feature extraction program 816. The feature extraction program 816 extracts information necessary for determining the important sentence from the morphological analysis data (FIG. 7—step 78), and passes the extracted information to the decision tree operation program 808. The decision tree operation program 808 accesses a decision tree database 814 prepared in advance and determines whether or not the sentence is a summary sentence based on the information extracted by the feature extraction program 816 (FIG. 7—Step 79). If it is a summary sentence, the important sentence display program 810 is executed, the display on the display means 812 is displayed in a highlighted color, otherwise it is displayed in the background color, and the process proceeds to the next sentence. After the summary sentence determination process is completed, the synopsis display program 811 is executed to display the first sentence of each paragraph of the displayed sentence in an emphasized color different from the important sentence (FIG. 7—Step 80).
[0047]
As described above, according to the present embodiment, for example, article search according to a search condition such as a keyword and rapid reading support can be processed as one flow.
[0048]
Example 3
FIG. 11 is a configuration diagram of an example in which the fast reading support method described in the second embodiment is realized in the embodiment of the network type document search support service. In FIG. 11, it is assumed that a service providing apparatus (server) and a service receiver apparatus (client 1 and client 2 are connected via an information communication network. For this reason, the server will be described with reference to FIG. The system bus 804 is provided with a communication unit 1101 and an interface 1102 with an information communication network, for the sake of simplicity, the display of other devices is omitted for the server. An output means, here means a printer, etc. 1122 is a CPU, and executes processing according to a program described later, 1123 is an input means, for example, a keyboard, a mouse, etc. 1112 is a system bus. Reference numeral 1113 denotes display means, which is a so-called display such as CTP. 1114 is a search / speed reading support interface program holding means, for example, a hard disk is used, 1115 is a work area of a memory, 1116 is a communication means, 1111 is an interface, and client 1 In this example, only the client 1 is specifically illustrated as having the same configuration, and the client 2 is simplified as a display of only the bus 1132 and the interface 1131.
[0049]
First, the user inputs a document search service use start request command through the input unit 1123. Then, the request command is transmitted to the server side by the communication means 1116 through the communication network. The server that has received the command transmits the search / speed reading support interface operation program 805 stored in the program holding means to the client 1 via the communication network. When the client 1 receives the program 805, it holds it in the search / speed reading support interface program holding means 1114 and operates the program using the calculation resources (CPU 1122, work area 1115). Then, a screen for requesting a search keyword appears as described in FIG. The user inputs a search keyword through the input unit 1123. The input keyword is transmitted to the server side by the communication means 271116. Then, the server operates the search program 805 and starts searching based on the transmitted keyword. Next, the obtained search result is transmitted to the client 1 via the communication network. The client 1 presents the result to the user using the interface operation program 805 transmitted and held. The display contents at this time are the same as those shown in FIG. When the user selects a text display button here, a text display request is transmitted to the server side via the network, and the server sends a corresponding document to the client 1 in response to the request, and operates on the computer of the client 1. The interface operation program 805 displays the sent document on the screen. The display contents at this time are the same as those shown in FIG. However, at this time, a copy of the sent document is left in the work area on the server side. When the user further turns on the speed reading support switch 1001, the request is sent to the server via the network, and the server receives the request and sends the sent document remaining in the work area to the speed reading support program. In accordance with the same procedure as shown in the second embodiment, the speed reading support program performs morphological analysis, feature extraction, important sentence determination by the decision tree operation program, and finally synopsis extraction for each sentence in the document. The server sends information such as which sentence is displayed in which color to the client 1 via the network. As a result, the interface operation program of the client 1 can adjust the display color of each sentence in the document.
[0050]
Since this operation is the same for the client 2, the description thereof is omitted.
[0051]
Example 4
In the third embodiment, the search / speed reading support interface operation program 805 is transmitted from the server to the client 1 via the communication network in response to the search request from the client. The same operation as in the third embodiment can be performed. Also in this case, as described with reference to FIG. 11, when the user inputs a document search request command using the input unit 1123, the interface operation program stored in the program holding unit 1114 on the client 1 is started, The operation procedure is processed in the same manner as in the third embodiment to enable search from a remote place and fast reading support.
[0052]
【The invention's effect】
As is clear from the above description, according to the present invention, since the summary sentence is displayed as it is in the text, the context before and after the summary sentence is preserved and the summary is easy to read. Furthermore, by displaying the first sentence of each paragraph together with the summary, the outline of the contents can be understood.
[0053]
When a decision tree for each genre is accumulated in advance and referred to, a summary sentence determination specialized for a specific genre can be performed very effectively.
[Brief description of the drawings]
FIG. 1 is a block diagram showing the concept of data processing of a speed reading support method according to the present invention.
FIG. 2 is a diagram showing an example of an article adopted as a target of fast reading support.
FIG. 3 is a diagram illustrating a result of performing feature extraction on each sentence of the article in FIG.
FIG. 4 is a diagram showing an example of a decision tree
FIG. 5 is a flowchart illustrating the concept of data processing in FIG. 1;
6 is a view showing a display example of a result of speed reading support for the article of FIG. 2;
FIG. 7 is a block diagram showing the concept of data processing when the document speed reading support method according to the present invention is applied to a document search support apparatus.
FIG. 8 is a diagram showing an example of a device configuration that realizes the processing shown in FIG. 7;
FIG. 9 is a diagram showing a specific example of a display form of a document search result.
FIG. 10 is a diagram showing a specific example when a specific document body is displayed according to a document search result.
FIG. 11 is a diagram showing a specific example of an embodiment for receiving a document search support service to which a speed reading support method according to the present invention is applied from a remote place;

Claims (1)

A means to input text consisting of a headline and text;
Means for holding information about the genre of the text,
Information about a genre is acquired based on a predetermined rule for the input sentence, and a feature analysis is performed for each sentence in the sentence according to the predetermined rule while referring to a decision tree corresponding to the genre. Means to determine,
Depending on the result of determining whether or not the sentence is a summary sentence, in the case of a summary sentence, it is displayed in a highlighted color, otherwise it is displayed in a color different from the highlighted color, and the first sentence of each paragraph of the input sentence Means for displaying in a color different from the two colors,
The feature analysis is performed by coding the position in the sentence, the similarity to the headline, the TFIDF in the sentence, the number of characters in the sentence, and the position in the paragraph for each sentence.
The means for determining whether or not the summary sentence is:
(1) Determine the similarity between each sentence and the headline,
(2-1) If the similarity is less than or equal to the first similarity, determine a position in the sentence;
(3-1) If the position in the sentence is equal to or less than the first predetermined value, it is determined as a “non-summary sentence”;
(3-2) If the position in the sentence is greater than the first predetermined value, determine the position in the paragraph;
(4-1) If the position in the paragraph is equal to or less than the second predetermined value, it is determined as a “summary sentence”;
(4-2) If the position in the paragraph is larger than the second predetermined value, the position in the sentence is determined again,
(5-1) If the position in the sentence is equal to or less than a third predetermined value, it is determined as a “non-summary sentence”;
(5-2) If the position in the sentence is greater than the third predetermined value, determine the number of characters in the sentence;
(6-1) If the number of characters in the sentence is equal to or less than a fourth predetermined value, it is determined as a “non-summary sentence”;
(6-2) If the number of characters in the sentence is greater than a fourth predetermined value, it is determined as a “summary sentence”;
(2-2) If the similarity is greater than the first similarity, the similarity is determined again,
(7-1) If the similarity is greater than the second similarity, determine “summary sentence”;
(7-2) If the similarity is equal to or less than the second similarity, TFIDF is determined,
(8-1) If the TFIDF is greater than the fifth predetermined value, determine “non-summary sentence”;
(8-2) If the TFIDF is less than or equal to the fifth predetermined value, determine the number of characters in the sentence;
(9-1) If the number of characters in the sentence is greater than a sixth predetermined value, determine “summary sentence”;
(9-2) If the number of characters in the sentence is less than or equal to the sixth predetermined value, determine TFIDF again,
(10-1) If the TFIDF is equal to or less than a seventh predetermined value, it is determined as a “summary sentence”,
(10-2) A speed reading support device characterized in that if the TFIDF is larger than a seventh predetermined value, it is determined as a “non-summary sentence”.

Images